Glomeruli Detection And Classification Algorithm In Continuous Slice Images Based On Improved Faster R-CNN And CBP Methods

Objective: The detection of microstructures in continuous tissue sections of kidneys is of great significance for the study of basic nephrology.With the development of deep learning and computer-assisted technology,the identification of microstructures of kidneys and the development of renal microscopy can through the object detection algorithm in computer vision.The classification of the glomeruli development period can provide a means for the morphological study of the development of kidney and the development of glomeruli.Methods: In this study,three consecutive slices of kidney development in mice were selected,including 14 days of embryonic age,17 days of embryonic age and 5 postnatal days.A total of 265 images were labeled,and 18,098 glomeruli were labeled as our study data set.The data distribution method was used to alleviate the uneven distribution of data in the four stages of glomerular development.In order to solve the task of identifying glomeruli in kidney tissue slice images,this experiment uses a Faster R-CNN based deep learning recognition method and uses ResNet-50-FPN as a basic feature extractor to achieve glomeruli detection and classification in mouse kidneys of three developmental stages.In order to improve the performance of the glomerular detection model in two-dimensional slices,this study proposes a complementary candidate box algorithm.Based on the characteristics of two-dimensional sequential images,a virtual three-dimensional target space is established and the result of the object is corrected.Results: This study trained and tested multiple data sets,including the original data set and its three subsets of data during the kidney development period,and the augmented data set and its three subsets of data during the kidney development period.The ratio of training and testing data set is 8:2.By comparing the performance of multiple models,the performance of the model trained by augmented data is improved by 33%,the accuracy is 0.91,and the recall is 0.90.In addition,after the correction of the complementary candidate box algorithm proposed in this study,the model accuracy is 0.94 and the recall is 0.92.Conclusion: This study realized the detection and staging of glomeruli in images of mouse kidney tissue slices from three developmental stages by deep learning methods,providing an automated detection method for basic research in nephrology,reducing human and material resources consumption and improving the efficiency of basic nephrology research,will help the development and progress of future nephrogenesis research.